OpenMP Canonical Loop Operation for OMP Dialect (WIP)

[abidmalikwaterloo]

Summary

This is a Work In Progress.
The patch define the OpenMP Canonical Operation for OpenMP MLIR Dialect

[abidmalikwaterloo]

mlir/include/mlir/Dialect/OpenMP/OpenMPOps.td

470

Do we need LoopControlCanonical for canonical operation? The current is Fortran-style and may not exactly match the naive C/C++ for-stmt.

[kiranchandramohan]

mlir/include/mlir/Dialect/OpenMP/OpenMPOps.td

462

Depends on what is the value that is returned. What value are you expecting to return here?

473

Try to reuse the same loop control unless you need a different one.

You might also want to discuss whether the loop should unit step forward counting as in the OpenMP IRBuilder or general like in the current LoopControl.

[abidmalikwaterloo]

mlir/include/mlir/Dialect/OpenMP/OpenMPOps.td

462

@Meinersbur @kiranchandramohan My understanding is that the return result is an array of loop features that are OpenMP canonicalized. We can have ArrayAttr. We need Variadic because loop nesting can be variable.

473

@Meinersbur Please see the comments for the increment.
@kiranchandramohan We need a conditional variable to take care of C/C++

for (init-expr; test-expr; incr-expr) structured-block

We can add an optional condition variable for this format in the LoopControl.

[Meinersbur]

mlir/include/mlir/Dialect/OpenMP/OpenMPOps.td

438

473

Personally, I would only make it take the trip count N with normalized %iv = 0...N-1 and let the Frontend handle all language semantics (is ub inclusive?, negative step, integer overflow, non-integer iteration variables, ...). It also simplifies mid-end passes since they only have to handle this normalized form.

However, for omp.wsloop the Fortran people had strong opinion that we should represent lb/ub/step explicitly. IMHO it only makes the MLIR representation more complex. Apart from edge cases (lb=INT_MIN, ub=INT_MAX) it more expressive than the normalized way, so I am willing to compromise.

[kiranchandramohan]

@abidmalikwaterloo What is the plan here? Would @do or @Meinersbur be taking over this patch? Do you have a timeline?

[abidmalikwaterloo]

In D147658#4477588, @kiranchandramohan wrote:

@abidmalikwaterloo What is the plan here? Would @do or @Meinersbur be taking over this patch? Do you have a timeline?

@kiranchandramohan I am working on this with the help of @Meinersbur . I am aiming to wind up this ASAP.

mlir/include/mlir/Dialect/OpenMP/OpenMPOps.td

oilist(`linear` `(`
                 $schedule_chunk_var, type($schedule_chunk_var)) `)`
           |`nowait` $nowait
           |`ordered` `(` $ordered_val `)`
           |`order` `(` custom<ClauseAttr>($order_val) `)`
           |`reduction` `(`
               custom<ReductionVarList>(
                 $reduction_vars, type($reduction_vars), $reductions
               ) `)`
      ) `for` custom<LoopControl>($region, $lowerBound, $upperBound, $step,
                                   type($step), $inclusive) attr-dict
   }];
   let hasVerifier = 1;
 }
 
+//===---------------------------------------------------------------------===//
+// OpenMP Canonical Loop Operation              
+//===---------------------------------------------------------------------===//
+def CanonicalLoopOp : OpenMP_Op<"canonical_loop", [AttrSizedOperandSegments,
+                     AllTypesMatch<["lowerBound", "upperBound", "step"]>]> {
+ let summary = "OpenMP Canonical Loop Operation"; 
+  let description = [{  
+    This is a work in progress. The aim of the operation is to return
+    a <loop type> that confirms the canonical form as describted by the 
+    OpenMP Specification https://www.openmp.org/spec-html/5.0/openmpsu40.html. 
+    The lower and upper bounds specify a half-open range: the range includes 
+    the lower bound but does not include the upper bound. If the `inclusive` 
+    attribute is specified then the upper bound is also included. For the 
+    following FORTRAN code:
+    ```
+    DO u = lb1, ub1
+      dosomthing (u)
+    DO v = lb2, ub2
+         dosomthing (v)
+    END DO
+    END DO
+    ```
+   
+   ``` 
+   %outerinfo, %innerinfo = omp.canonical_loop for (%iv1) : 
+      i32 =  (%lb1) to (%ub1) step (%step1) {
+            do_somthing (%iv1)
+            %inner = omp.canonical_loop for (%iv2) : 
+              i32 =  (%lb2) to (%ub2) step (%step2) {
+                   do_somthing (%iv2)
+                  omp.yield
+             }     
+    } 
+       omp.yield %innerinfo
+    }
+    ```                    
+     "(%lb) to (%ub) step (%step)" is Fortran-style and may not exactly 
+     match the naive C/C++ for-stmt. Optional If_expr/conditional check can
+     be used for the C/C++ style for loop 
+                       }];
+          let arguments = (ins Variadic<IntLikeType>:$lowerBound,
+             Variadic<IntLikeType>:$upperBound,
+             Variadic<IntLikeType>:$step,
+             Optional<I1>:$if_expr,
+             UnitAttr:$inclusive
+     );
+   
+     // let results = (outs Variadic<AnyType>:$loopInfo);
+     // Do we need to define a type for this? 
+     // that will contain all the information 
+     // about the canonical loop
+     let regions = (region AnyRegion:$region);
+     
+     
+     
+      let assemblyFormat = [{
+      oilist(`if` `(` $if_expr `)`)
+     `for` custom<LoopControl>($region, $lowerBound, $upperBound, $step, 
+                                  type($step), $inclusive) attr-dict 
+  }];
+    // Should we use the same LoopControl or
+    // define a new one for canonical loop operation
+    //
+     
+     let hasCustomAssemblyFormat = 1;
+    
+}
+
 //===----------------------------------------------------------------------===//
 // Simd construct [2.9.3.1]
 //===----------------------------------------------------------------------===//
 
 def SimdLoopOp : OpenMP_Op<"simdloop", [AttrSizedOperandSegments,
                          AllTypesMatch<["lowerBound", "upperBound", "step"]>]> {
  let summary = "simd loop construct";
   let description = [{